Private AI governance workflow
Logo Proposal Validation Agent
An internal AI-assisted brand governance workflow for logo proposals, designed around deterministic policy checks, retrieved guidance, structured outputs, reviewer-ready responses, and human review.
Role
AI Product Architect / Technical Lead
University of Arizona
Tags
AI Review WorkflowHuman-in-the-loopPolicy Retrieval
Problem
Logo and brand proposal approvals were consuming reviewer time because submissions needed policy checks, judgment, drafted replies, and follow-up for unclear cases.
Users
Brand reviewers, marketing stakeholders, campus submitters, and leadership teams interested in faster governed review workflows.
Ownership
- Owned project inception, architecture, development, publishing, documentation, and stakeholder coordination.
- Designed the AI review flow from submission intake through approve, deny, ambiguous review request, and clarifying follow-up.
- Coordinated across development and stakeholder groups as interest grew beyond the original reviewer.
Hard Parts
- Fine-tuned retrieval quality and data quality so RAG-backed policy guidance became more accurate.
- Discovered the request form itself had unclear fields and one question people often answered strategically just to move through the form faster.
- Used findings from the AI workflow to help improve the underlying request process, not just automate review.
Leadership
- Made the technical architecture decisions and drove the project from conversation to shipped workflow.
- Wrote documentation and coordinated with stakeholders as the workflow expanded.
- Kept policy-sensitive decisions human-in-the-loop while increasing confidence in routine cases.
Shipped
- AI review agent with approve, deny, ambiguous, and clarification paths.
- RAG and deterministic-policy checks with structured outputs.
- Channel notifications with approval and denial choices.
- Clarifying-question flow for ambiguous submissions.
- Tool calling to aid decisions as the system matured.
Impact
- Reduced manual review burden and increased stakeholder confidence through iteration.
- Helped reviewers focus on submissions that actually needed human judgment.
- Created a reusable governed-review pattern for brand and policy workflows.
Signals
Typical reviewer volume: roughly six submissions per day.Manual review, reply drafting, and outreach previously took about 10 minutes per submission.Clear cases can now move through the workflow in under a minute.
Highlights
- Mapped the real review problem into approve, deny, clarify, and escalate paths for policy-sensitive decisions.
- Combined deterministic rules, RAG guidance, LLM classification, and structured outputs.
- Planned reviewer-ready response drafting and Microsoft Teams Adaptive Cards workflow concepts for adoption inside existing operations.
- Kept ambiguous or policy-sensitive cases human-in-the-loop.
- Treated the engagement as a reusable pattern for policy retrieval, AI triage, and governed internal review workflows.
Tools
PythonFastAPIOpenAI APIRAGEmbeddingspgvectorPostgresStructured JSONDeterministic rulesHuman reviewEscalation pathsMicrosoft Teams Adaptive Cards